| 9 Dec 2025 |
 SomeoneSerge (back on matrix) | Robbie Buxton: have you seen https://linus.schreibt.jetzt/posts/ubuntu-images.html? Also stumbled on https://github.com/numtide/nix-vm-test while trying to remember the url | 16:35:16 |
| SomeoneSerge (back on matrix) | Nix in this case is just a tool we already had for memoization and isolation, it just happens to come with a particularly shitty scheduler | 16:36:28 |
 Robbie Buxton | No I haven’t, I’ll take a look, cheers! | 16:37:26 |
| SomeoneSerge (back on matrix) | Any other memoization solution one could build must necessarily be a version of Nix (on top of the actual Nix), likely stripped down and using heuristics instead of trying to faithfully compute the hash "of that and only that which matters"... | 16:38:47 |
| SomeoneSerge (back on matrix) | Which with the current state of affairs is probably cheaper and more efficient 🤷 | 16:39:14 |
| SomeoneSerge (back on matrix) | But mathematically still the same thing | 16:39:27 |
 tomberek | Anyone using impure derivations to explicitly have non-caching behavior? | 16:42:29 |
| SomeoneSerge (back on matrix) | No, that's not really useful? | 17:07:37 |
| 10 Dec 2025 |
| SomeoneSerge (back on matrix) | * No, that's not really useful?
Edit: we generally do want memoization (caching), as well as ability to deterministically name things in sentences "X succeeds (fails)". Just that so far we haven't included all relevant platform info in the derivation, so the equivalence classes are too big. It's one line fixable though
| 18:29:41 |
| SomeoneSerge (back on matrix) | * No, that's not really useful?
Edit: we generally do want memoization (caching), as well as ability to deterministically name things in sentences "X succeeds (fails)". Just that so far we haven't included all relevant platform info in the derivation, so the equivalence classes are too big. It's one line fixable though
Impure derivations just add sandboxing that needs to be bypassed, and a scheduler that is more of a liability
| 18:30:29 |
 Ari Lotter | hm looks like vllm fails to build because outlines has a test that imports libcuda.so on collection.. | 22:03:19 |
| SomeoneSerge (back on matrix) | https://github.com/NixOS/nixpkgs/pull/465751#issuecomment-3604113652 | 23:30:45 |
 apyh | hahaha you're already ahead of me! | 23:31:20 |
| SomeoneSerge (back on matrix) | Gaetan has been on it | 23:32:00 |
| apyh | yeah, seems like there's a clear path to patch / upstream a fix to llama cpp python to make it lazy | 23:33:22 |
| apyh | btw i posted logs about that torch nvrtc thing | 23:33:36 |
| 11 Dec 2025 |
 Gaétan Lepage | Yes, but it's not as straightforward as they initialize some top-level constants at module import time with the content of the loaded library...
I still think that it would be the best solution to this issue. | 09:19:45 |
|  youthlic changed their profile picture. | 14:36:35 |
